Test the impact before you make the change.
GitHub tells you whether your code is safe to merge. Preflight tells you whether the change is safe for your company — and it answers in revenue, customers and commitments rather than in dependencies.
ChangeAI_MODEL· Model provider
Route the research agent to a cheaper model provider
OpenAIDeepSeek
- DriftChecking…Pending
- MarginChecking…Pending
- ReleaseChecking…Pending
- ContractChecking…Pending
- ProofChecking…Pending
- ClaimChecking…Pending
Every system in your company answers a different question. None of them answer yours.
What you can already see
- Code review
- Is the diff correct?
- CI
- Do the tests pass?
- Analytics
- What happened last week?
- Finance
- What did we spend?
What nothing tells you
- Is this change safe for the business?
- Who is worse off if we ship it?
- What did we already promise that this breaks?
Those three questions get answered by a meeting, from memory, by whoever happens to be in the room. Preflight answers them from the graph, with the evidence attached, before the change is made.
Five changes that have nothing in common. One way to answer them.
A model swap is an engineering decision. A price cut is a finance decision. Removing a feature is a product decision. Preflight treats all three as the same question.
Proposed changes
ChangeAI_MODEL· Model provider
Route the research agent to a cheaper model provider
OpenAIDeepSeek
- DriftChecking…Pending
- MarginChecking…Pending
- ReleaseChecking…Pending
- ContractChecking…Pending
- ProofChecking…Pending
- ClaimChecking…Pending
- ARR at stake
- $481K
- Customers
- 284
- Margin
- +3.8 pts
- Contracts
- 2
- Claims
- 1
- Engineering
- ~3.5 days
- Risk
- High
- Projected saving
- +$18.4K / yr
- Accounts re-routed
- 412
- Workflows affected
- 17
- Regressions found
- 38
Preflight is the surface. These are what it asks.
Each system reads a different part of the Company Graph and contributes one kind of consequence. A Preflight calls whichever of them the change actually touches.
- MarginFinancial consequenceCore
- DriftDependency riskCore
- ProofCommitmentsCore
- ReleaseBehavioural regressionsBeta
- ContractContractual exposureBeta
- ClaimPublic claimsBeta
Can propose a Preflight
One Preflight, open in the workspace.
A proposal, the systems that answered it, what it does to the company, and the safer version of the same change. The figures are the ones from the homepage — it is the same Preflight.
ChangeAI_MODEL· Model provider
Route the research agent to a cheaper model provider
OpenAIDeepSeek
- Projected saving
- +$18.4K / yr
- ARR exposed
- $481K
- Customers affected
- 284accounts
- Failed checks
- 1of 6 checks
- DriftTarget API stable · no deprecation openPass
- Margin+$18,400/year across 412 accountsPass
- Release38 regressions across 17 workflowsWarning
- Contract2 controls name the current providerWarning
- Proof1 open commitment assumes EU inferenceWarning
- Claim“EU-hosted inference” no longer holdsFail
Route EU accounts to an EU-hosted model, everything else to DeepSeek
- 0 contracts affected
- 0 claims invalidated
- Saving re-measured when run
- ARR at stake
- $481K
- Customers
- 284
- Margin
- +3.8 pts
- Contracts
- 2
- Claims
- 1
- Engineering
- ~3.5 days
- Risk
- High
Read this column before the one on the left. The checks say what happens to the software; these seven rows say what happens to the company.
Most tools stop at the dependency. Nogu keeps going.
The same traversal, followed until it reaches something you can decide about. Five layers of the company model, one proposed change, and every relationship it touches on the way down.
- Dependency
openai-node
v4.28
- Service
research-agent
3 call sites
- Infrastructure
eu-central-1
inference
- Feature
Research Agent
- Workflow
Deep research
17 affected
- Release
v4.3
38 regressions
- Plan
Enterprise
- Subscription
Annual
284 active
- Customer
284 accounts
12 at renewal
- Cost
Model inference
−$1,533 / mo
- Revenue
Attributed ARR
$481K
- Margin
Affected revenue
+3.8 pts
- Promise
EU inference
1 open
- Contract
Data residency
2 controls
- Claim
EU-hosted inference
invalidated
A dependency graph stops at the first band. The decision lives in the last three.
Where a blast radius ends
A dependency graph can tell you that a package is used in three services and that one of them runs in Frankfurt. That is true, and it is the first band of five.
It cannot tell you that the workflow those services power is the one 284 enterprise accounts renew on, that $481K of ARR sits behind them, or that two signed controls and one public claim name the provider you are about to remove.
Those are not engineering facts. They are the reason the change gets made or does not.
From a proposal to a plan.
- 01
Propose
Describe the change — a provider swap, a price, a feature you want gone. Or let a pull request, a Sunset candidate or your own API propose it.
- 02
Fan out
Preflight works out which systems the change touches and asks each one in parallel.
- 03
Translate
Findings come back as revenue, customers, contracts and effort. Never as a dependency count.
- 04
Decide
Safe, review or block — then a plan: the fixes Nogu can run, the risks you choose to accept, and a check on the pull request.
What a Preflight is designed to reach
Stripe, GitHub, PostHog, HubSpot can be connected today. Staging, Your website work without a connector. The rest are sources the model is built for and cannot be connected yet. The app never offers a provider whose consent flow it cannot complete.
Money
- Stripe (available today)
- Chargebee
- Paddle
- AWS Cost Explorer
Product & code
- GitHub (available today)
- GitLab
- PostHog (available today)
- Amplitude
- Staging (available today)
Commitments
- HubSpot (available today)
- Salesforce
- DocuSign
- Your website (available today)
Run one against a change you are already arguing about.
Connect three sources and Preflight will answer the decision that is currently sitting in a thread. If it tells you nothing you did not know, that is a real answer and we would like to hear it.